Matrix and Tensor Factorization Methods for Natural Language

Matrix and Tensor Factorization Methods for
Natural Language Processing
Guillaume Bouchard∗ Jason Naradowsky# Sebastian Riedel#
Tim Rockt¨aschel# and Andreas Vlachos#
∗
Xerox Research Centre Europe
[email protected]
#
Computer Science Department
University College London
{j.narad, s.riedel, t.rocktaschel, a.vlachos}@cs.ucl.ac.uk
1
Tutorial Objectives
Tensor and matrix factorization methods have attracted a lot of attention recently thanks to their
successful applications to information extraction,
knowledge base population, lexical semantics and
dependency parsing. In the first part, we will first
cover the basics of matrix and tensor factorization
theory and optimization, and then proceed to more
advanced topics involving convex surrogates and
alternative losses. In the second part we will discuss recent NLP applications of these methods and
show the connections with other popular methods
such as transductive learning, topic models and
neural networks. The aim of this tutorial is to
present in detail applied factorization methods, as
well as to introduce more recently proposed methods that are likely to be useful to NLP applications.
2
2.1
Tutorial Overview
Matrix/Tensor Factorization Basics
In this part, we first remind essential results on
bilinear forms, spectral representations of matrices and low-rank approximation theorems, which
are often omitted in undergraduate linear algebra
courses. This includes the link between eigenvalue decomposition and singular value decomposition and the trace-norm (a.k.a. nuclear norm) as
a convex surrogate of the low-rank constraint on
optimization problems. Then, an overview of the
most efficient algorithms to solve low-rank constrained problems is made, from the power iteration method, the Lanczos algorithm and the implicitly restarted Arnoldi method that is implemented in the LAPACK library (Anderson et al.,
1999). We show how to interpret low-rank models
as probabilistic models (Bishop, 1999) and how
we can extend SVD algorithms that can factor-
ize non-standard matrices (i.e. with non-Gaussian
noise and missing data) using gradient descent, reweighted SVD or Frank-Wolfe algorithms. We
then show that combining different convex objectives can be a powerful tool, and we illustrate it by
deriving the robust PCA algorithm by adding an
L1 penalty term in the objective function (Cand`es
and Recht, 2009). Furthermore, we introduce
Bayesian Personalized Ranking (BPR) for matrix
and tensor factorization which deals with implicit
feedback in ranking tasks (Rendle et al., 2009). Finally, will introduce the collective matrix factorization model (Singh and Gordon, 2008) and tensor extensions (Nickel et al., 2011) for relational
learning.
2.2
Applications in NLP
In this part we will discuss recent work applying matrix/tensor factorization methods in the context of NLP. We will review the Universal Schema
paradigm for knowledge base construction (Riedel
et al., 2013) which relies on matrix factorization and BPR, as well as recent extensions of
the RESCAL tensor factorization (Nickel et al.,
2011) approach and methods of injecting logic
into the embeddings learned (Rockt¨aschel et al.,
2015). These applications will motivate the connections between matrix factorization and transductive learning (Goldberg et al., 2010), as well
as tensor factorization and multi-task learning
(Romera-Paredes et al., 2013). Furthermore, we
will review work on applying matrix and tensor
factorization to sparsity reduction in syntactic dependency parsing (Lei et al., 2014) and word representation learning (Pennington et al., 2014). In
addition, we will discuss the connections between
matrix factorization, latent semantic analysis and
topic modeling (Stevens et al., 2012).
3
Structure
Part I: Matrix/Tensor Factorization Basics (90
minutes)
• Matrix factorization basics (40 min): bilinear forms, spectral representations, low rank
approximations theorems, optimization with
stochastic gradient descent, losses
• Tensor factorization basics (20 minutes):
representations,notation
decompositions
(Tucker etc.)
• Advanced topics (30 minutes): convex surrogates, L1 regularization, alternative losses
(ranking loss, logistic loss)
Break (15 minutes)
Part II: Applications in NLP (75 minutes)
• Information extraction, knowledge base population with connections to transductive
learning and multitask learning (35 minutes)
• Lexical semantics with connections to neural
networks, latent semantic analysis and topic
models (30 minutes)
• Structured prediction (10 minutes)
4
About the Speakers
Guillaume Bouchard is a senior researcher in
statistics and machine learning at Xerox, focusing
on statistical learning using low-rank model for
large relational databases. His research includes
text understanding, user modeling, and social media analytics. The theoretical part of his work is
related to the efficient algorithms to compute high
dimensional integrals, essential to deal with uncertainty (missing and noisy data, latent variable
models, Bayesian inference). The main application areas of his work includes the design of virtual conversational agents, link prediction (predictive algorithms for relational data), social media
monitoring and transportation analytics. His web
page is available at www.xrce.xerox.com/
people/bouchard.
Jason Naradowsky is a postdoc at the Machine
Reading group at UCL. Having previously obtained a PhD at UMass Amherst under the supervision of David Smith and Mark Johnson, his current
research aims to improve natural language understanding by performing task-specific training of
word representations and parsing models. He is
also interested in semi-supervised learning, joint
inference, and semantic parsing. His web page is
available at http://narad.github.io/.
Sebastian Riedel is a senior lecturer at University College London and an Allen Distinguished
Investigator, leading the Machine Reading Lab.
Before, he was a postdoc and research scientist
with Andrew McCallum at UMass Amherst, a researcher at Tokyo University and DBCLS with
Tsujii Junichi, and a PhD student with Ewan Klein
at the University of Edinburgh. He is interested
in teaching machines how to read and works at
the intersection of Natural Language Processing
(NLP) and Machine Learning, investigating various stages of the NLP pipeline, in particular those
that require structured prediction, as well as fully
probabilistic architectures of end-to-end reading
and reasoning systems. Recently he became interested in new ways to represent textual knowledge
using low-rank embeddings and how to reason
with such representations. His web page is available at http://www.riedelcastro.org/.
Tim Rockt¨aschel is a PhD student in Sebastian Riedel’s Machine Reading group at University College London. Before that he worked as
research assistant in the Knowledge Management
in Bioinformatics group at Humboldt-Universit¨at
zu Berlin, where he also obtained his Diploma
in Computer Science. He is broadly interested
in representation learning (e.g. matrix/tensor factorization, deep learning) for NLP and automated
knowledge base completion, and how these methods can take advantage of symbolic background
knowledge. His webpage is available at http:
//rockt.github.io/.
Andreas Vlachos is postdoc at the Machine
Reading group at UCL working with Sebastian
Riedel on automated fact-checking using lowrank factorization methods. Before that he was
a postdoc at the Natural Language and Information Processing group at the University of
Cambridge and at the University of WisconsinMadison. He is broadly interested in natural language understanding (e.g. information extraction,
semantic parsing) and in machine learning approaches that would help us towards this goal.
He has also worked on active learning, clustering and biomedical text mining. His web page
is available at http://sites.google.com/
site/andreasvlachos/.
References
[Anderson et al.1999] Edward Anderson, Zhaojun Bai,
Christian Bischof, Susan Blackford, James Demmel,
Jack Dongarra, Jeremy Du Croz, Anne Greenbaum,
S Hammerling, Alan McKenney, et al. 1999. LAPACK Users’ guide, volume 9. SIAM.
[Bishop1999] Christopher M Bishop. 1999. Bayesian
PCA. In Advances in Neural Information Processing Systems, pages 382–388.
[Cand`es and Recht2009] Emmanuel J Cand`es and Benjamin Recht. 2009. Exact matrix completion via
convex optimization. Foundations of Computational
mathematics, 9(6):717–772.
[Goldberg et al.2010] Andrew Goldberg, Ben Recht,
Junming Xu, Robert Nowak, and Xiaojin Zhu.
2010. Transduction with matrix completion: Three
birds with one stone. In Advances in Neural Information Processing Systems 23, pages 757–765.
[Lei et al.2014] Tao Lei, Yu Xin, Yuan Zhang, Regina
Barzilay, and Tommi Jaakkola. 2014. Low-rank
tensors for scoring dependency structures. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1381–
1391.
[Nickel et al.2011] Maximilian Nickel, Volker Tresp,
and Hans-Peter Kriegel. 2011. A three-way model
for collective learning on multi-relational data. In
Proceedings of the 28th International Conference on
Machine Learning (ICML-11), pages 809–816.
[Pennington et al.2014] Jeffrey Pennington, Richard
Socher, and Christopher D. Manning. 2014. Glove:
Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods
in Natural Language Processing.
[Rendle et al.2009] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme.
2009. Bpr: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth
Conference on Uncertainty in Artificial Intelligence,
pages 452–461.
[Riedel et al.2013] Sebastian Riedel, Limin Yao, Benjamin M. Marlin, and Andrew McCallum. 2013.
Relation extraction with matrix factorization and
universal schemas. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Atlanta, GA.
[Rockt¨aschel et al.2015] Tim Rockt¨aschel, Sameer
Singh, and Sebastian Riedel. 2015. Injecting
Logical Background Knowledge into Embeddings
for Relation Extraction . In Proceedings of the
2015 Human Language Technology Conference of
the North American Chapter of the Association of
Computational Linguistics.
[Romera-Paredes et al.2013] Bernardino
RomeraParedes, Hane Aung, Nadia Bianchi-Berthouze,
and Massimiliano Pontil.
2013.
Multilinear
multitask learning. In Proceedings of the 30th
International Conference on Machine Learning,
pages 1444–1452.
[Singh and Gordon2008] Ajit P. Singh and Geoffrey J.
Gordon. 2008. Relational learning via collective
matrix factorization. In Proceedings of the 14th
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 650–658.
[Stevens et al.2012] Keith Stevens, Philip Kegelmeyer,
David Andrzejewski, and David Buttler. 2012. Exploring topic coherence over many models and many
topics. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language
Learning, pages 952–961.